Dragon versus Nakamura

[Nordlandia]

One way to justify more handicap is by playing FRC. In Fischer Random,the human is out of the comfort zone and has to extemporize from scratch. Based on my testing, Dragon is nearly on par Stockfish in FRC. So Dragon is not necessarily weaker in Fischer Random.

[lkaufman]

One way to justify more handicap is by playing FRC. In Fischer Random,the human is out of the comfort zone and has to extemporize from scratch. Based on my testing, Dragon is nearly on par Stockfish in FRC. So Dragon is not necessarily weaker in Fischer Random.

Yes, CCRL has already confirmed our own findings that Dragon is stronger than SF12 in FRC. We did indeed demonstrate that it is possible even for Komodo MCTS (pre-Dragon) to give knight odds in FRC Rapid to a fairly strong GM (Alex Lenderman), even though in normal chess Komodo needed some compensation for the knight (a pawn or castling rights). Perhaps Dragon MCTS could now give knight odds in FRC Rapid to any human with decent chances. The ideal opponent would be Wesley So as he is still the reigning FRC World Champion.

[Chessqueen]

One way to justify more handicap is by playing FRC. In Fischer Random,the human is out of the comfort zone and has to extemporize from scratch. Based on my testing, Dragon is nearly on par Stockfish in FRC. So Dragon is not necessarily weaker in Fischer Random.

Yes, CCRL has already confirmed our own findings that Dragon is stronger than SF12 in FRC. We did indeed demonstrate that it is possible even for Komodo MCTS (pre-Dragon) to give knight odds in FRC Rapid to a fairly strong GM (Alex Lenderman), even though in normal chess Komodo needed some compensation for the knight (a pawn or castling rights). Perhaps Dragon MCTS could now give knight odds in FRC Rapid to any human with decent chances. The ideal opponent would be Wesley So as he is still the reigning FRC World Champion.

What a Great Idea! GM SO is also playing LIVE Now in the Skilling Open ==> https://en.chessbase.com/post/skilling-open-2020-live

[Chessqueen]

One way to justify more handicap is by playing FRC. In Fischer Random,the human is out of the comfort zone and has to extemporize from scratch. Based on my testing, Dragon is nearly on par Stockfish in FRC. So Dragon is not necessarily weaker in Fischer Random.

Yes, CCRL has already confirmed our own findings that Dragon is stronger than SF12 in FRC. We did indeed demonstrate that it is possible even for Komodo MCTS (pre-Dragon) to give knight odds in FRC Rapid to a fairly strong GM (Alex Lenderman), even though in normal chess Komodo needed some compensation for the knight (a pawn or castling rights). Perhaps Dragon MCTS could now give knight odds in FRC Rapid to any human with decent chances. The ideal opponent would be Wesley So as he is still the reigning FRC World Champion.

What a Great Idea! GM SO is also playing LIVE Now in the Skilling Open, can you imagine all the headlines around the World Komodo Beat the Current FRC Champion giving him a Knight Odds ! But you will have to find a good sponsor, since GM SO is NOT going to play cheap to risk being embarrassed like this.

[jp]

It shouldn't actually be very complicated to build a search function that in each position selects objectively sub-optimal move that, however, leads to positions where all but one or very few moves for opponent are bad.

That would be a very interesting and useful function to have for chess analysis (not just so an engine can beat up a human in handicap matches).

I've certainly proposed it. Team K/D seems to indicate K/D in MCTS mode already 'sort of does this'...but an engine that would basically say 'option 2' gives the opponent 75% (or a number) of 'reasonable' looking moves that lead to an even worse continuation, would be ideal. Developers are only interested in elo for best lines though it seems.

Also, it would require a special GUI I think...and 'engine developers' specialize in tweaking parameters and such (or organizing odds exhibitions) not coding GUI's, so don't expect it to ever happen.

Not really the proper thread for this specific idea. But to follow up, I would add that Komodo/Dragon using their MCTS...in conjunction with Chessbase would be ideal since about everyone has Chessbase and us it to work on their game, prepare opening idea, etc. A boy can dream...

For a start at least, I think we could get away without a new GUI. We could just specify this mode in the engine (say Komodo/Dragon) and it could just use the existing eval. field in the GUI to give the output.

... The idea is interesting though, but not worded properly, since choosing moves where most replies are bad typically means choosing captures, and trading off is a bad strategy when behind in material. But looking for moves that appear to offer many plausible replies, most of which can be refuted with deeper search, is a good idea. I think MCTS does this to some extent, but it could be improved on.

Yes, it seems that it requires a bit of thought about how to implement it. Ideally you want it to be able to judge not just how many good and bad replies there are, but how difficult they are to find. You could start by making capture/re-captures an exception in the code.

MCTS is not good enough for what we want. MCTS will at best approximate what we want in an accidental and uncontrolled way. For handicap matches, the handicap-giver is forced to bluff, but in prep/analysis you are not looking to bluff. e.g. for White opening preparation you are looking for awkward lines for the unbooked opponent among all the +/= lines.

[Cornfed]

It shouldn't actually be very complicated to build a search function that in each position selects objectively sub-optimal move that, however, leads to positions where all but one or very few moves for opponent are bad.

That would be a very interesting and useful function to have for chess analysis (not just so an engine can beat up a human in handicap matches).

I've certainly proposed it. Team K/D seems to indicate K/D in MCTS mode already 'sort of does this'...but an engine that would basically say 'option 2' gives the opponent 75% (or a number) of 'reasonable' looking moves that lead to an even worse continuation, would be ideal. Developers are only interested in elo for best lines though it seems.

Also, it would require a special GUI I think...and 'engine developers' specialize in tweaking parameters and such (or organizing odds exhibitions) not coding GUI's, so don't expect it to ever happen.

Not really the proper thread for this specific idea. But to follow up, I would add that Komodo/Dragon using their MCTS...in conjunction with Chessbase would be ideal since about everyone has Chessbase and us it to work on their game, prepare opening idea, etc. A boy can dream...

For a start at least, I think we could get away without a new GUI. We could just specify this mode in the engine (say Komodo/Dragon) and it could just use the existing eval. field in the GUI to give the output.

... The idea is interesting though, but not worded properly, since choosing moves where most replies are bad typically means choosing captures, and trading off is a bad strategy when behind in material. But looking for moves that appear to offer many plausible replies, most of which can be refuted with deeper search, is a good idea. I think MCTS does this to some extent, but it could be improved on.

Yes, it seems that it requires a bit of thought about how to implement it. Ideally you want it to be able to judge not just how many good and bad replies there are, but how difficult they are to find. You could start by making capture/re-captures an exception in the code.

MCTS is not good enough for what we want. MCTS will at best approximate what we want in an accidental and uncontrolled way. For handicap matches, the handicap-giver is forced to bluff, but in prep/analysis you are not looking to bluff.[/b] e.g. for White opening preparation you are looking for awkward lines for the unbooked opponent among all the +/= lines.

Actually, the point above I put in 'red' just now is not always the case when it comes to prep/analysis. Lets say you are playing a truly speculative opening/defense - The Scandanvian Portuguese variation for example, you do not always want to play (as black) the 'best computer moves' simply because there are lines for white that are easy to find and keep/increase his edge with ease.

What you want to be able to do in that case if find...2nd or 3rd best moves which are not too far off the reservation, but offer white MORE chances to go wrong (not talking just outright blunders here) in the rough and tumble of actual play...particularly OTB, especially at the increasingly common quicker time controls.

[carldaman]

I agree, the ability and opportunity to bluff should be a real consideration when prepping for human opponents. An engine that can go for bluffing lines is in an asset in my 'book'.

[jp]

For handicap matches, the handicap-giver is forced to bluff, but in prep/analysis you are not looking to bluff. e.g. for White opening preparation you are looking for awkward lines for the unbooked opponent among all the +/= lines.

Actually, the point above I put in 'red' just now is not always the case when it comes to prep/analysis. Lets say you are playing a truly speculative opening/defense - The Scandanvian Portuguese variation for example, you do not always want to play (as black) the 'best computer moves' simply because there are lines for white that are easy to find and keep/increase his edge with ease.

What you want to be able to do in that case if find...2nd or 3rd best moves which are not too far off the reservation, but offer white MORE chances to go wrong (not talking just outright blunders here) in the rough and tumble of actual play...particularly OTB, especially at the increasingly common quicker time controls.

It's a matter of what we mean by "bluff". If you choose e.g. a really unsound gambit opening, you could say you've already bluffed by playing that opening. Then, further down the opening, choosing a variation within it that has comp eval e.g. -0.8 instead of a line with comp eval -0.65 is not what I'd call bluffing. (It seems this is your "not too far off the reservation" move.)

The other type of bluffing is not really knowing the eval or lines and just playing/evaluating moves semi-blind. That's not the type of bluffing I'd want the computer to give me, because I'd want to know what sort of risk playing that line involves.

[Chessqueen]

One way to justify more handicap is by playing FRC. In Fischer Random,the human is out of the comfort zone and has to extemporize from scratch. Based on my testing, Dragon is nearly on par Stockfish in FRC. So Dragon is not necessarily weaker in Fischer Random.

Yes, CCRL has already confirmed our own findings that Dragon is stronger than SF12 in FRC. We did indeed demonstrate that it is possible even for Komodo MCTS (pre-Dragon) to give knight odds in FRC Rapid to a fairly strong GM (Alex Lenderman), even though in normal chess Komodo needed some compensation for the knight (a pawn or castling rights). Perhaps Dragon MCTS could now give knight odds in FRC Rapid to any human with decent chances. The ideal opponent would be Wesley So as he is still the reigning FRC World Champion.

What a Great Idea! GM SO is also playing LIVE Now in the Skilling Open, can you imagine all the headlines around the World Komodo Beat the Current FRC Champion giving him a Knight Odds ! But you will have to find a good sponsor, since GM SO is NOT going to play cheap to risk being embarrassed like this.

If any of the these previous cham[ions would be alive, Capablanca, Alekhine, or Fischer which of these would do better against Komodo MCTS with the same 2 pawns Odds that Nakamura lost 4 straight games the second day they played?

[lkaufman]

One way to justify more handicap is by playing FRC. In Fischer Random,the human is out of the comfort zone and has to extemporize from scratch. Based on my testing, Dragon is nearly on par Stockfish in FRC. So Dragon is not necessarily weaker in Fischer Random.

Yes, CCRL has already confirmed our own findings that Dragon is stronger than SF12 in FRC. We did indeed demonstrate that it is possible even for Komodo MCTS (pre-Dragon) to give knight odds in FRC Rapid to a fairly strong GM (Alex Lenderman), even though in normal chess Komodo needed some compensation for the knight (a pawn or castling rights). Perhaps Dragon MCTS could now give knight odds in FRC Rapid to any human with decent chances. The ideal opponent would be Wesley So as he is still the reigning FRC World Champion.

What a Great Idea! GM SO is also playing LIVE Now in the Skilling Open, can you imagine all the headlines around the World Komodo Beat the Current FRC Champion giving him a Knight Odds ! But you will have to find a good sponsor, since GM SO is NOT going to play cheap to risk being embarrassed like this.

If any of the these previous cham[ions would be alive, Capablanca, Alekhine, or Fischer which of these would do better against Komodo MCTS with the same 2 pawns Odds that Nakamura lost 4 straight games the second day they played?

Well, the obvious answer would be Capablanca, since he was known for technique and almost never losing, while Alekhine took lots of risks. But the level of play advanced over time, so Fischer would be my pick, and probably Carlsen today would be the best of all time at playing odds vs. engines. But I think that the time limit is too fast for any human to win a match from Dragon at the two White pawn handicap used here. I've learned a lot from this match and from the comments about how to simulate the play of top human players in situations like this; in my opinion the simulations overestimated the strength of Nakamura (or any human) at this time limit. I think this was the bigger factor in the result rather than the MCTS choice; in most critical positions regular Dragon would choose the same move. Humans just don't play very well with short time.